"Got infinite logprob" assertion failure, with option condition_on_previous_text=False

[ItakeLs]

Sometimes running audio triggers the "got infinite logprob" assertion, all audio that triggers this does work in the whisper model from the OpenAI repo. The error occurs in the "may_flush_segment" function

# see GreedyDecoder.update()
chunck_indices = chunk_tokens_nosot + [tokenizer.eot]
assert len(chunk_logprobs) == len(chunck_indices), f"{len(chunk_logprobs)} != {len(chunck_indices)}"
logprobs = [logprob[i] for (logprob, i) in zip(chunk_logprobs, chunck_indices)]
assert min([p.isfinite().item() for p in logprobs]), "Got infinite logprob"

A sample of audio that I could get to reliably reproduce this error was the mp4 from this youtube link -> https://www.youtube.com/watch?v=D9G1VOjN_84 I downloaded the MP4 from here -> https://yt1ss.net/en?q=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DD9G1VOjN_84 (would upload but 10mpbs limit, filesize is 18mb)

The audio was run on the medium model size, with condition_on_previous_text=False and the remaining parameters untouched

[Jeronymous]

Thanks for reporting and reporting a way to reproduce. That seems to be due to "condition_on_previous_text=False". I'll fix it

[seanth]

Here to report the same bug.